In this work, we explored the feasibility of predicting the structural drift from the first seconds of P-wave signals for On-site Earthquake Early Warning (EEW) applications. To this purpose, we investigated the performance of both linear least square regression (LSR) and four non-linear machine learning (ML) models: Random Forest, Gradient Boosting, Support Vector Machines and K-Nearest Neighbors. Furthermore, we also explore the applicability of the models calibrated for a region to another one. The LSR and ML models are calibrated and validated using a dataset of ~6000 waveforms recorded within 34 Japanese structures with three different type of construction (steel, reinforced concrete, and steel-reinforced concrete), and a smaller one of data recorded at US buildings (69 buildings, 240 waveforms). As EEW information, we considered three P-wave parameters (the peak displacement, Pd, the integral of squared velocity, IV2, and displacement, ID2) using three time-windows (i.e., 1, 2, and 3 seconds), for a total of nine features to predict the drift ratio as structural response. The Japanese dataset is used to calibrate the LSR and ML models and to study their capability to predict the structural drift. We explored different subsets of the Japanese dataset (i.e., one building, one single type of construction, the entire dataset. We found that the variability of both ground motion and buildings response can affect the drift predictions robustness. In particular, the predictions accuracy worsens with the complexity of the dataset in terms of building and event variability. Our results show that ML techniques perform better than LSR models, likely due to the complex connections between features and the natural non-linearity of the data. Furthermore, we show that by implementing a residuals analysis, the main sources of drift variability can be identified. Finally, the models trained on the Japanese dataset are applied the US dataset. We found that the exporting EEW models worsen the prediction variability, but also that by including correction terms as function of the magnitude can mitigate such problem. In other words, our results show that the drift for US buildings can be predicted by minor tweaks to models.

中文翻译：

---

使用机器学习回归器进行结构漂移预测的地震早期预警系统

在这项工作中，我们探索了从 P 波信号的前几秒预测结构漂移用于现场地震早期预警 (EEW) 应用的可行性。为此，我们研究了线性最小二乘回归 (LSR) 和四种非线性机器学习 (ML) 模型的性能：随机森林、梯度提升、支持向量机和 K-最近邻。此外，我们还探索了针对一个区域校准的模型对另一个区域的适用性。LSR 和 ML 模型使用在 34 个日本结构中记录的约 6000 个波形的数据集进行校准和验证，这些结构具有三种不同的结构（钢、钢筋混凝土和钢筋混凝土），以及在美国建筑物中记录的较小数据（69 个建筑物，240 个波形）。作为 EEW 信息，我们使用三个时间窗口（即 1、2 和 3 秒）考虑了三个 P 波参数（峰值位移 Pd、平方速度的积分 IV2 和位移 ID2），总共有九个特征预测漂移比作为结构响应。日本数据集用于校准 LSR 和 ML 模型并研究它们预测结构漂移的能力。我们探索了日本数据集的不同子集（即，一座建筑物，一种单一类型的建筑，整个数据集。我们发现地面运动和建筑物响应的可变性都会影响漂移预测的鲁棒性。特别是，预测精度恶化随着数据集在构建和事件可变性方面的复杂性。我们的结果表明 ML 技术比 LSR 模型表现更好，可能是由于特征之间的复杂联系和数据的自然非线性。此外，我们表明，通过实施残差分析，可以确定漂移变异性的主要来源。最后，将在日本数据集上训练的模型应用于美国数据集。我们发现导出的 EEW 模型会恶化预测的可变性，但通过包含作为幅度函数的校正项可以缓解此类问题。换句话说，我们的结果表明，可以通过对模型进行细微调整来预测美国建筑物的漂移。我们发现导出的 EEW 模型会恶化预测的可变性，但通过包含作为幅度函数的校正项可以缓解此类问题。换句话说，我们的结果表明，可以通过对模型进行细微调整来预测美国建筑物的漂移。我们发现导出的 EEW 模型会恶化预测的可变性，但通过包含作为幅度函数的校正项可以缓解此类问题。换句话说，我们的结果表明，可以通过对模型进行细微调整来预测美国建筑物的漂移。